switch constant issue

Mon Jun 9 17:55:00 GMT 2003

>>>>> "Andrew" == Andrew Haley <aph@redhat.com> writes:

Andrew> I guess this is OK, but I'm not sure. It does rather break
Andrew> -fno-assume-compiled in that the constant uses early binding, so we
Andrew> might up with a different value for a constant that the one in its
Andrew> definition.
We just have to follow the Java language and binary compatibility
rules. In this case, inlining the constant is correct, and in fact
required.
Basically, this is a problem with Java for which the spec just
mandates a particular solution. It isn't perfect; the standard class
libraries themselves change incompatibly from time to time :-(. For
instance there is a constant in java.awt.event that changed in recent
memory (I forget which one offhand).
Jeff, I think my patch won't affect Steve's problem. That patch was
for a different case, where we would inline a non-{int,string}
constant incorrectly (causing link errors, as I recall). Steve is
seeing the opposite: we aren't inlining an int constant that we
should. I suspect this is a different front end problem (perhaps
unrelated to -fno-assume-compile -- I think we may have a PR for this
already, but I haven't looked).
Steve, your install directory says "3.4" but you are really using 3.3.
-fno-assume-compiled isn't supported on any branch yet, but there are
needed bug fixes in 3.4 that we won't be putting into the 3.3 series.
Tom